Re: [TLS] supported_versions question

[Matt Caswell <[e-mail]>]

On 31 October 2016 at 23:13, David Benjamin <[e-mail]> wrote:
> On Mon, Oct 31, 2016 at 6:34 PM Kurt Roeckx <[e-mail]> wrote:
>>
>> On Mon, Oct 31, 2016 at 07:11:10PM +0000, David Benjamin wrote:
>> > On Mon, Oct 31, 2016 at 2:57 PM Ilari Liusvaara
>> > <[e-mail]>
>> > wrote:
>> >
>> > > On Mon, Oct 31, 2016 at 06:43:52PM +0000, Matt Caswell wrote:
>> > > > A few supported_versions questions:
>> > > >
>> > > > 1) What should a server do if supported_versions is received but
>> > > > ClientHello.legacy_version != TLS1.2? Fail the handshake, or just
>> > > > ignore legacy_version?
>> > >
>> > > If legacy_version > TLS1.2, the spec requires server to ignore
>> > > legacy_version.
>> > >
>> > > The case where legacy_version < TLS1.2 IIRC isn't specified, but
>> > > ignoring legacy_version is reasonable in this case too.
>> > >
>> > > > 2) What should a server do if supported_versions is received,
>> > > > ClientHello.legacy_version == TLS1.2, but supported_versions does
>> > > > not
>> > > > contain TLS1.3 or TLS1.2 (e.g. it contains TLS1.1 or below)? Fail
>> > > > the
>> > > > handshake, use the legacy_version, or use use the versions in
>> > > > supported_versions?
>> > >
>> > > There's also the case where supported_versions has TLS 1.1 and TLS
>> > > 1.4,
>> > > the latter the server has never heard about...
>> > >
>> > > > 3) If the answer to (2) above is ignore the legacy_version, and just
>> > > > use the versions in supported_versions, which client_version should
>> > > > be
>> > > > used in the RSA pre-master secret calculation? The one in
>> > > > legacy_version, or the highest one in supported_versions? Presumably
>> > > > it has to be the one in legacy_version, otherwise thing will fail
>> > > > when
>> > > > the client talks to a server that doesn't understand
>> > > > supported_versions?
>> > >
>> > > Yeah, I presume putting the version in legacy_version is the only sane
>> > > thing to do. But causes other problems with downgrade protection.
>> > >
>> > > OTOH, RSA key exchange is known to be very broken and is affected by
>> > > all kinds of downgrade (and other) attacks. So if one wants actual
>> > > security, it needs to be removed.
>> > >
>> >
>> > We could say the versions extension only applies to 1.2 and up. I.e.
>> > don't
>> > bother advertising 1.1 and 1.0 as a client and servers ignore 1.1 and
>> > 1.0
>> > when they see them in the version list. That keeps the protocol
>> > deployable
>> > on the Internet as it exists, avoids having to evaluate too versioning
>> > schemes (if you see the extension, you don't bother reading
>> > legacy_version
>> > at all), while avoiding the weird behavior where, given this
>> > ClientHello:
>> >
>> >    legacy_version: TLS 1.2
>> >    supported_versions: {TLS 1.1}
>> >
>> > TLS 1.3 says to negotiate TLS 1.1 and TLS 1.2 says to negotiate TLS 1.2.
>>
>> So I guess you're also saying that a server that implements TLS
>> 1.1 to TLS 1.3, but disables TLS 1.2 and TLS 1.3 support should
>> ignore the supported_versions even when it knows about it?
>>
>> I guess I have same questions but with only TLS 1.3 disabled, to
>> be sure when we need to look at it.
>
>
> Hrm, actually I hadn't thought of that. Yeah, I guess a server which
> disables versions must then gate supported_version handling on whether TLS
> 1.3 is enabled. That's not a dealbreaker, but is certainly additional
> gnarliness.
>
> (Our current implementation just processes the extension uncondtionally, but
> we'll also happily negotiate old versions out of it.)

I came up with some alternative wording that I think captures the discussion:

https://github.com/tlswg/tls13-spec/pull/748

Matt